Research

Competitive landscapes, market scans, external grounding. What exists outside Lawhive, read closely and mapped to what we build.

• A Benchmark Factory for Lawrence
  2026-08-26 · Why no published legal benchmark fits Lawhive's need to prove Lawrence in each newly acquired jurisdiction and practice area, and how the benchmark factory generates tasks on demand: a synthetic matter file in the shape of the platform's own virtual file system, a task a lawyer would set, and criteria grounded verbatim in the generated facts and retrieved law. Walks the eight generation stages with excerpts from a real generated bundle, and the five principles behind the design. Draft.
• Spike Report: Passive Time Capture
  2026-08-21 · Can a lawyer's billable time be recorded without the lawyer recording it? A browser recording is reduced to a compact "activity stream" naming what was worked on, on which matter, and for how long; a language model turns the stream into proposed time entries with cited evidence. Hypothesis, method, results across a multi-matter multi-tab session, the ABA 93-379 framing of what may bound a proposed duration, and what the activity stream unlocks beyond billing. Final.
• The Revolut operating model, as one picture — a mental model for the Ways of Working group
  2026-08-18 · The QuantumLight playbooks (104 articles) reduced to one picture: a goals loop (company → department → team KPIs, SQL-first behind an 80% quality gate, read weekly by exception), a people loop (role = 3–4 skills, scorecards, a talent bar per seniority, proficiency-vs-bar grades with calibration), and the spine that keeps both honest (one platform fed from the warehouse, plus a Performance Team reporting to the CEO). Maps each of the working group’s five decision areas onto the picture. Pre-read for Kha-ai and Matt. Draft.
• Grok Bot, read closely — what xAI shipped and what Lawrence should take from it
  2026-08-17 · The landing page says each Bot gets its own cloud computer; xAI’s own docs say all your Bots share one and tell you not to treat a Bot as a security boundary. Every claim read first-hand from x.ai/bot and 12 doc pages, plus the launch chatter. What actually ships (computer use works; no audit log, no dry run, no undo, nothing written about prompt injection), the four patterns worth taking into Lawrence (skill vs routine, the manual-to-automated ladder, a pause that asks a human for one input, policy written at the moment of the decision), and the six to reject. Draft for discussion.
• Ledgers & reconciliation — a plain-language explainer
  2026-08-06 · A jargon-free tour of double-entry accounting, ledgers, and reconciliation, drawn from the Monzo reconciliation team: the accounting equation, cash vs accruals, what a “break” is, and the timezone and netting traps behind a £1.5M unexplained cash position that once stopped us closing the financial year. Ends with concerns about the V3 cashiering & reconciliation design, chiefly the world account and how it is journaled. Draft for discussion.
• How fast can chat reveal text without hitting the stream head? — pacing mode benchmark
  2026-07-17 · Drove both chat backends against a real matter, recorded six token-arrival traces, and simulated four pacing modes over them. Pacing is a jitter-buffer problem with two axes that trade off: lag vs smoothness. Today’s prod pacing runs ~10s behind on a real summary; fixed-fast is timely but burst-freezes ~460ms; adaptive is the only mode both timely and smooth, and its shipped horizon is too loose — tuning to ~one chunk-gap halves the lag (~1.4s→~0.7s). Also corrects an earlier claim: both backends stream the answer text at the same ~400ms cadence. Live demo replaying real traces + the tuning recommendation.
• Migrating case context off the dying GPT-5 models — per-stage benchmark
  2026-07-14 · The December migration for our highest-volume LLM pipeline (~167k gens/30d), benchmarked by replaying real production generations from Langfuse traces. 3,120 replays + 930 two-judge verdicts. Per-stage picks: luna for the minis, terra@medium for CCO generation, gpt-5.5 for reconciliation (terra objectively disqualified — 32–58% of its reconcile ops fail to apply), materiality deferred, 8 dead prompts to delete. Final.
• Which model should write ingestion’s summaries? — GPT-5.6 migration benchmark
  2026-07-12 · The prod summarisation model (gpt-5-mini@low) is deprecated on Dec 11; we benchmarked it against gpt-5.4-mini and the GPT-5.6 family (sol/terra/luna × reasoning efforts) on 39 real documents through the exact production code path. 718 generations, 1,014 gpt-5.5 judgements, latency/cost/hallucination analysis, and a verbosity-bias correction on the pairwise reads. Outcome: gpt-5.6-luna@medium as a parameter-only change, terra@medium as the quality-first alternative. Final.
• Making @lawhive/ui agent-ready — design systems for AI coding agents
  2026-07-03 · The 2025-26 playbook for making a design system readable and usable by AI coding agents, adversarially verified (23 confirmed / 2 refuted claims) and mapped onto packages/ui. Six mechanisms (machine-readable manifests, MCP servers, llms.txt, pit-of-success APIs, closed token contracts, guardrails), the honest evidence-quality flags, a priority-ordered gap list, and a foundations-first sequence. Research synthesis, draft for discussion.
• The shareability gap — Quick by Shopify & the internal artifact platform landscape
  2026-06-12 · Close read of Shopify's Quick (53,069 internal sites in 11 months on one $200 VM): architecture, capability APIs, the four decisions that made it work, and the growth inflection when agents got good. Plus a scan of the open-source and vendor landscape (Smallweb, Val Town, Netlify Drop, tiiny.host, surge.sh, Coolify, Vercel preview comments, AI artifact surfaces) on a publish-friction × audience-trust map, the honest Quick ↔ html-previews capability comparison, and the 8-step roadmap in dependency order. Draft for discussion.